Can Edge-Deployable Vision-Language Models Identify Species?
For edge vision, specialized training outweighed model size: 300M-parameter BioCLIP beat 2–8B VLMs, while field imagery degraded every model and open-set prompts produced invented species.
Four **2–8B VLMs** and the 300M-parameter BioCLIP were tested on 96 species using clean photos and camera-trap images. BioCLIP led every VLM by **33.2–59.2 percentage points** on the expanded sample.
For an offline vision agent, choose domain-matched training before adding parameters, and evaluate on images captured by the intended hardware. Constrain or validate open-set labels against a taxonomy.
Four **2–8B VLMs** and the 300M-parameter BioCLIP were tested on 96 species using clean photos and camera-trap images. BioCLIP led every VLM by **33.2–59.2 percentage points** on the expanded sample. For an offline vision agent, choose domain-matched training before adding parameters, and evaluate on images captured by the intended hardware. Constrain or validate open-set labels against a taxonomy. Every model lost **9.6–26.6 points** on field imagery, suggesting image legibility is a shared bottleneck. Open prompting also produced nonexistent species in **5.9–9.6%** of responses; the study does not establish performance outside this task.
This confirms that domain-matched training can matter more than parameter count for a narrow vision task, while narrowing evaluation requirements to deployment-realistic imagery and taxonomy-valid outputs. The large field-image drop shows that clean-photo results do not transfer intact to camera-trap conditions, and invented species labels make open-set generation an additional reliability problem rather than merely a classification error.